Take the ordinary disjoint union, and then add a new _|_ element, distinct from both existing copies of _|_ (which are still distinct from each other).
Now why is that not the category-theoretic coproduct? h . Left = f and h . Right = g both for _|_ and for finite elements of the types. And it looks universal to me.
Yeah, but there could be more functions from Either X Y to Z than pairs of functions from X to Z and from Y to Z. For example, if z :: Z, then you have two functions h1 and h2 such that h1 . Left = const z and h1 . Right = const z and the same holds for h2. Namely, h1 = const z h2 = either (const z) (const z) This functions are different : h1 (_|_) = z while h2 (_|_) = (_|_). And if Either X Y was a category-theoretic coproduct, then the function from Either X Y to Z would be UNIQUELY determined by it's restrictions to X and Y.